Researchers at the AI Accountability Lab (AIAL) have released new findings addressing the lack of visibility into AI training data despite it being an obligation under the AI Act. Funded by Mozilla and set to be presented at the ACM Conference on Fairness, Accountability, and Transparency (FAccT) in June 2026, the research highlights a neglected provision in the European Union’s AI Act requiring General-Purpose AI providers to publish a summary of how their models were trained.
The work shows that despite nearly a year since its inception, only 4 providers have published their summaries, with all of them being small and/or open-source models. As enforcement is set to kick in from August later this year, the availability of the public summary is a crucial resource for stakeholders to defend and exercise their rights. The work provides a quality assessment framework to evaluate and assist in ensuring the information is sufficiently transparent and useful that can be used by providers to self-improve their summaries as well as by the European Commission’s AI Office in assessing whether summaries are complete and produced in good faith. The paper and the associated resources can be found at: https://aial.ie/research/gpai-training-transparency/
The AIAL, founded and led by Dr Abeba Birhane, focuses on the societal impact of AI technologies. The lab aims to foster an AI ecosystem that prioritises public interest, especially the most marginalised and disenfranchised, from research and product development to regulatory oversight.
